Research on semantic segmentation algorithm of remote sensing images based on improved SegNet

DOI:	10.11809/bqzbgcxb2023.07.031
Keywords:	remote sensing image; semantic segmentation; SegNet; deep learning; image processing
Abstract:	Semantic segmentation of remote sensing images is widely used in production and life. Aiming at the problem that SegNet, a commonly used semantic segmentation network, is not accurate enough for small target and edge detail segmentation on remote sensing images, this paper proposes an EP SegNet model to improve SegNet. This model replaces the activation function of ReLU with ELU to speed up convergence and avoid neuron death. It also removes the pooling of the last layer of the encoder for convolution only to reduce the loss of spatial information, builds a Bottleneck layer to deepen the network while reducing the number of parameters, and introduces pyramid pooling module (PPM) to improve network awareness of the global information. The experimental results show that the accuracy of EP SegNet is 97.48%, which is 3.31% higher than that of SegNet. Meanwhile, segmentation accuracy of the edge details and the multi scale targets improves significantly in image segmentation, which shows that the model performs better than the original model does.
Published:	2023-07-28
PDF HTML